isprs annals IV 2 W4 303 2017

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

PARAMETER ESTIMATION AND MODEL SELECTION FOR INDOOR
ENVIRONMENTS BASED ON SPARSE OBSERVATIONS
Y. Dehbia,∗ , S. Loch-Dehbia,∗ and L. Plümera
a

Institute of Geodesy and Geoinformation, University of Bonn, Meckenheimer Allee 172, Bonn, Germany (dehbi, loch-dehbi, pluemer)@igg.uni-bonn.de
Commission IV, WG IV/7

KEY WORDS: 3D building indoor models, CityGML, Gauss-Markov model, Model selection, Gaussian mixture, Symmetry

ABSTRACT:
This paper presents a novel method for the parameter estimation and model selection for the reconstruction of indoor environments
based on sparse observations. While most approaches for the reconstruction of indoor models rely on dense observations, we predict
scenes of the interior with high accuracy in the absence of indoor measurements. We use a model-based top-down approach and
incorporate strong but profound prior knowledge. The latter includes probability density functions for model parameters and sparse
observations such as room areas and the building footprint. The floorplan model is characterized by linear and bi-linear relations
with discrete and continuous parameters. We focus on the stochastic estimation of model parameters based on a topological model
derived by combinatorial reasoning in a first step. A Gauss-Markov model is applied for estimation and simulation of the model

parameters. Symmetries are represented and exploited during the estimation process. Background knowledge as well as observations
are incorporated in a maximum likelihood estimation and model selection is performed with AIC/BIC. The likelihood is also used for
the detection and correction of potential errors in the topological model. Estimation results are presented and discussed.
1 MOTIVATION AND CONTEXT
Indoor models are of great interest in a wide range of applications. They are of high relevance for indoor navigation, evacuation planning, facility management or guide for the blind. While
models of the exterior such as models in level of detail 3 (LoD3)
according to CityGML (Gröger and Plümer, 2012) are by now
state of the art, modelling of indoor environments is not yet widely
available and the acquisition of the corresponding data remains
expensive. While most approaches require measurements of high
density such as 3D point clouds or images, we propose an approach which gets along with few observations in order to predict
floorplans of high accuracy.
The challenge is to estimate an indoor model based on few observations. More precisely, the task that we consider in this paper
is to place a set of n rectangular rooms within a polygonal footprint, locating the doors and estimating the height of the rooms.
The resulting model is characterized by discrete and continuous
model parameters that are related by both linear and non-linear
constraints. Our approach is characterized by sparse observations such as room areas, functional use, window locations and,
possibly, room numbers that can be acquired from building management services. Indoor images or laser scans are not required.
A maximum a-posteriori (MAP) estimation uses further statistical knowledge such as probability density functions and Gaussian
mixtures for model parameters.

In this paper, we assume that the topological model is provided
by a preceding step using Constraint Logic Programming as presented by Loch-Dehbi et al. (2017). The topological model consists of the room neighbourhood information and the correspondence between rooms and windows. This article focuses on the
optimal estimation of model parameters. Therefore, we use a
Gauss-Markov model with Bayesian background knowledge. We
represent and exploit symmetries and repetitive structures. This
contributed to improving the existing approach and thus the estimation of model parameters. We asses the quality of the model by
a maximum likelihood estimation and compare different models

by the information criteria Bayesian information criterion (BIC)
and Akaike information criterion (AIC). This allows for the comparison of model hypotheses based on their likelihood.
The main contribution of this paper is the improved estimation of
model parameters based on a-priori derived topological models
performed by Loch-Dehbi et al. (2017). This includes:

• the use of prior knowledge in the form of probability density
functions for model parameters,
• an MAP-estimation considering observations such as the area
of rooms, the location of windows and the building footprint,
• the representation and exploitation of symmetries, especially
translational ones,

• detection and correction of errors in the topological model,
• calculation of normalized likelihoods for each model hypothesis and comparison of models with a derived ranking.

All in all, for the prediction and estimation of a floorplan and its
parameters with high accuracy, the areas of the rooms, the footprint and the location of windows together with a prior knowledge
in form of probability density functions are sufficient.
The remainder of this article is structured as follows. The next
section discusses related work. Section 3 introduces the necessary theoretical background of the used stochastic approach.
Section 4 gives insight into the derivation of topological models
using constraint satisfaction. Section 5 explains our stochastic
estimation reasoning in detail, while Section 6 discusses the experimental results. The article is summarized and concluded in
Section 7.

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
303

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

2 RELATED WORK
Indoor models are gaining more and more attention for many
tasks such as rescue management. Turner and Zakhor (2014)
generated building floorplans from laser range data based on a
triangulation of a 2D sampling of wall positions. Becker et al.
(2015) proposed a grammar-based approach for the reconstruction of 3D indoor models from 3D point clouds. Yue et al. (2012)
used shape grammars describing a building style together with
a footprint and a set of exterior features to determine the interior layout of a building. Khoshelham and Dı́az-Vilariño (2014)
applied also shape grammars for 3D indoor modelling by iteratively placing, connecting and merging cuboid shapes. However,
a pre-design of the grammar rules is required. Ochmann et al.
(2016) segment a point cloud into rooms and outside area for reconstructing a scene using an energy minimization in a labelling
problem. In contrast to our approach, all mentioned approaches
require dense observations such as 3D point clouds from laserscans or range cameras using mobile mapping systems that are
both cost and time expensive. Measurements are expensive because each single room has to be accessed.
Various works tried to overcome the measurement overhead by
using low cost sensors. To this end, Rosser et al. (2015) presented
a method for constructing as-built plans of residential building interiors. The prediction is based on mobile phone sensor data and
improved by incorporating hard and soft constraints. Pintore et
al. (2016) generate 2.5D indoor maps based on images acquired
by mobile devices. Diakité and Zlatanova (2016) used the low

cost Android tablet from Google’s Tango project for the acquisition of indoor building environments. However, it is not possible
to derive detailed indoor models due to artefacts and anomalies in
the sensor data. Furthermore, the derivation of models from measurements is difficult due to the masking of walls by furniture.
As a consequence, our main contribution is the parameter estimation and model selection for unknown substructures in buildings
such as floorplans based on only few observations like the area of
rooms and footprints. We make use of strong model assumption
supported by a profound background knowledge.
In our context, Rosser et al. (2017) demonstrated a two-staged
semi automatic data-driven estimation of 2D building interior floorplans based on limited prior knowledge. However, their approach
requires beforehand topology of rooms and the orientation of
one room. Compared to Rosser et al. (2017), in our approach
the topology is automatically provided using Constraint Logic
Programming. Apart from stochastic reasoning, our approach
draws upon ideas of constraint satisfaction. A constraint-based
approach which generates possible floorplans respecting a set of
geometric constraints is presented by Charman (1994). This approach does not address as-built models and, however, does not
consider probable configurations. An overview of works in indoor modelling and mapping is given by Gunduz et al. (2016).
3 GAUSS-MARKOV MODELLING OF FLOORPLANS
The estimation of the location and shape parameters of floorplans has been addressed as a reasoning process which combines
constraint propagation and a maximum a-posteriori estimation

based on special graphical models (Loch-Dehbi et al., 2017). The
performed MAP-estimation is based mainly on the background
knowledge about room location and shape parameters. In order
to acquire more accurate estimations, observations such as window locations have to be integrated in the model selection as well.
Furthermore, regularities characterizing man-made objects such
as buildings are beneficial and have to be exploited. This paper

focuses on the integration, modelling and assessing of the impact
of these aspects on the estimation process. Hence, we provide a
method improving the mentioned MAP-estimation.
To this aim, based on a predetermined topological floorplan model
consisting of the neighbourhood of rooms and their window correspondence, the task we solve can be addressed as a stochastic parameter estimation defined in the form of a Gauss-Markov
model as follows:
˜
l = l + ṽ = Ax̃ + a, ṽ ∼ M (0, Σll )

(1)

where l refers to an observation vector which can deviate from
the true values ˜

l of the observations with residuals ṽ. x̃ denotes
the true values of the unknowns, i.e model parameters, while A
stands for the design matrix mapping the model parameters onto
the observations. By uncertain observations, a distribution M
is introduced. Our task lies in finding estimates x
b of the true
model parameters x̃ from the observations l. In our context, the
parameters as well as the residuals are a-priori unknown which
leads to an underconstrained problem. This can be overcome by
requiring that the weighted sum of the squared residuals
Ω(x) = v T (x)Σ−1
ll v(x),

(2)

of the residuals v(x) = Ax + a − l has to be minimal as follows:
−1 T −1
x
b = argminx Ω(x) = (AT Σ−1
A Σll (l − a).

ll A)

(3)

In the case of a normal distribution M , this is equivalent to determining the maximum likelihood estimation of the parameters:


1
1
T −1
p(l | x) =
v(x)
Σ
v(x)
. (4)
exp
−
ll
2
(2π)N |Σll |1/2

N is the rank of the covariance matrix Σll . Since our problem
is also characterised by bi-linear dependencies such as it is the
case between the room areas as multiplication of room shape parameters, the modelling of non-linear constraints has to be taken
into account. In this context, the linear problem is expanded by a
non-linear part modelled by non-linear Gauss-Markov model as
follows:
l ∼ M (f (x̃), Σll ),
(5)
where the parameters are related by a twice differentiable function f . In this case, the goal is to find an estimation:
x
b = argminx (l − f (x))T Σ−1
ll (l − f (x)),

(6)

which can be solved by Taylorization in the usual way (Niemeier,
2008) starting from approximate values of the unknown model
parameters. For more details, the interested reader is referred to
Förstner and Wrobel (2016).
Given the mathematical model, we are now interested in choosing

one model among several alternatives taking some observations l
into account. In our context, the observations consist of the locations of windows, room areas and the footprint geometry. To this
aim, a Bayesian approach is used in order to select a model M
having the largest posterior probability P (M | l) out of several
models Mm :
c = argmaxm p(l | Mm )P (Mm ).
M

The model is then defined following the Akaike information criterion (AIC) (Akaike, 1974) according to equation (7).
MAIC = argmaxm − log p(l | Mm ) + Um .

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
304

(7)

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

Um denotes the number of the parameters xm of the model Mm
characterizing the model complexity. By normally distributed
data the first term in equation 7 can be written as:
− log p(l | Mm ) =

1 T −1
vb Σll vb.
2

(8)

The model Mm can be likewise determined using the Bayesian
information criterion (BIC) (Schwarz et al., 1978) augmenting
the term Um by the number N of the observations as following:
MBIC = argmaxm − log p(l | Mm ) +

1
Um log N.
2

(9)

4 CONSTRAINT PROPAGATION FOR TOPOLOGICAL
FLOORPLAN DERIVATION
This section gives an overview about the method for deriving
topological floorplan models as prerequisite for the subsequent
stochastic reasoning. To solve the presented non-linear problem
with discrete and continuous parameters, we have principally to
cope with an a-priori infinite continuous search space. However,
we proposed a method in which the search space is narrowed
using architectural constraints and browsed by intelligent search
strategies using domain knowledge.
We start with solving a discrete combinatorial problem that yields
a good initialization for the stochastic reasoning. The correspondence of a-priori known windows to their rooms is unknown.
The bilateral relations between rooms have to be determined as
well. This consists in the knowledge about vertical and horizontal neighbourhood of rooms leading to a topological model. In order to make the combinatorial problem of locating rooms within
the footprint feasible, we build upon background knowledge consisting in probability density functions (PDFs) of shape and location parameters learned from extensive training data. We applied Gaussian mixtures as a good approximation to model skew
symmetric or multi-modal distributions, e.g. PDFs. Using Expectation Maximization (McLachlan and Peel, 2000), a Gaussian
mixture of m components each weighted by its probability ωi for
each continuous parameter is estimated for the reasoning process:
m
X

ωi N (µi , σi2 ).

Figure 1: Topological floorplan model derived using constraint
propagation
and depth, are restricted by upper and lower bounds in addition
to shape parameters of the corresponding footprint.
Since the floorplan model can be described by constraints on discrete as well as continuous variables with associated domains, it
can be addressed as a constraint satisfaction problem. By incorporating consistency-enforcing algorithms and constraint propagation, existing constraints are tightened or new constraints are
derived to narrow the search space (Dechter, 2003).
Figure 1 shows one topological model as a result of our constraint
propagation using Constraint Logic Programming (Frühwirth and
Abdennadher, 2003). It yields the bilateral relations between
rooms, the correspondence of windows to rooms and the selection of single components of the Gaussian mixtures for the model
parameters. We ignored temporarily the walls and relaxed the
search problem in order to provide buffers for constraint propagation with discrete domains. Consequently, the predicted preliminary rooms do not fill the entire footprint space. Particularly,
the alignment of rooms along a corridor is not guaranteed. The
acquired topological model serves as input for the parameter estimation and model selection as presented in the following section.
For more details on this combinatorial step the reader is referred
to Loch-Dehbi et al. (2017).

(10)
5 PARAMETER ESTIMATION AND MODEL

i=1

SELECTION OF FLOORPLANS
This allows for using well established stochastic reasoning afterwards.
We exploit architectural regularities and design corresponding
constraints on model parameters to restrict the search space. For
instance, it is not allowed that walls cross windows. Besides, a
lower window for example is part of an ith room if the y-coordinate of the room’s lower left corner yi equals the width of the
outer wall and the following constraint is satisfied:
((xi ≤ wiLoclj,1 ) ∧ (wiLoclj,2 ≤ (xi + widthi ))),

(11)

where wiLoclj,1 and wiLoclj,2 are the location parameters of a
lower window placed in the jth room as shown in Figure 2. The
constraint 11 implies in this case j = i. The single components
of the Gaussian mixture define upper and lower bounds for the
according continuous parameter. For example, the model parameters of the room, that is its lower left corner (xi , yi ) and its width

This section describes our approach for parameter estimation and
model selection for indoor models based on sparse observations.
Based on preliminary topological models acquired from a reasoning process following the method described in Section 4, an
optimal parameter estimation using a Gauss-Markov model and
Bayesian model selection is performed. The detailed method is
described by Loch-Dehbi et al. (2017). We exploit the symmetries, especially translational ones, characterising buildings in order to improve the estimation of model parameters. In this sense,
the distances between windows and walls as well as distances
between windows themselves are considered. These kinds of
symmetries are modelled in an explicit way. Models including
symmetries and those without symmetries are taken into consideration and compared as alternative models.
Our approach starts with defining the complete Gauss-Markov
model which consists of the functional and stochastic model including the design matrix A according to Equation 1. In our sce-

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
305

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

the according component of each parameter from these distributions in a preceding step. Thus, in this paragraph we can assume that model parameters are normally distributed with known
means µ and variances σ. We will use pseudo-observations for
the following model parameters:
• room widths widthi and depths depthi ,
• corridor width widthcorr and depth depthcorr ,
• outer and inner wall widths waOut and waIn.

Figure 2: Illustration of the used model parameters and observations.
nario, the additive term a is equal to zero. The model components
are elaborated in the following.
Functional model:
We assume that the adjacencies between the rooms and the respective windows as well as the ordering among the rooms are
given by the topological model as described above in Section 4.
The exact location and shape parameters are not required. In this
case, the functional model is uniquely specified by the following
model parameters:
• the widths and the depths of the lower (widthli , depthli ) and
upper (widthui , depthui ) rooms. The index i refers to the
ith room. As we assume that the rooms are aligned along
corridors, we have two single parameters (depthl , depthu )
both for the depth of the lower and the upper rooms, but
distinct parameters for the width of each room.
• two parameters for the width and the depth of the corridor
(widthcorr , depthcorr ).
• two parameters for the width of the walls. Thereby we assume that the widths of all outer walls (waOut) are the
same, as are the inner walls (waIn).
• two parameters for the width and the depth of the floorplan
which is assumed to be rectangular and, above, having its
lower left corner in the origin of the local coordinate system.
Thus they are named xmax and ymax .
• The area areai for each ith room and for the corridor.
Although it is not necessary to identify a model in a unique way,
we introduce the following parameters in order to relate the model
to the given observations:
• the location of the lower (wiLocli,1 , wiLocli,2 ) and upper
(wiLocui,1 , wiLocui,2 ) embrasures of the windows in the ith
room. We need only those embrasures which are directly
adjacent to a wall.
Figure 2 gives an overview of the domain model and its parameters. Following a Bayesian approach, we explicitly incorporate
prior knowledge and represent it by pseudo-observations following the ideas of Förstner and Wrobel (2016) and Bishop (2007).
Note that prior knowledge on model parameters is represented by
Gaussian mixture distributions and that we have already selected

Design matrix:
We apply the Gauss-Markov model where each observation is a
linear function of the model parameters. Since the Gauss-Markov
model is generic and the relation between model (parameters) and
observations is explicit, it can be used both for simulations and
parameter estimation. Thus, it is appropriate for the purpose of
this article.
For the design matrix we have linear and bi-linear constraints.
We start with the linear constraints. For a left embrasure in an ith
lower room, we have constraints which relate this observation to
the left adjacent wall:

waOut + (i − 1)waIn +

i−1
X

widthlj + wiDistli,1 = wiLocli,1

j=1

(12)
Likewise, we have for the right embrasures constraints which relate them to the right adjacent wall:

waOut + (i − 1)waIn +

i
X

widthlj − wiDistli,2 = wiLocli,2

j=1

(13)
The constraints for the embrasures in the upper rooms are formulated in an analogous way. The depth of the floorplan ymax is
related to the depth of the upper and lower rooms including the
corridor as well as the width of the respective walls. Since we
assume that rooms are aligned, we have a single depth parameter
for the upper and the lower row of rooms respectively:
2waOut + 2waIn + depthl + depthu + depthcorr = ymax
(14)
For the sequences of upper and lower rooms, we have two constraints which adds up their widths together with the widths of
the adjacent walls. For instance, the constraint for lower rooms is
formulated as follows:
2waOut + (r − 1)waIn +

r
X

widthli = xmax ,

(15)

i=1

where r refers to the room number in a room sequence. A similar
constraint holds for the corridor and its adjacent walls:
2waOut + widthcorr = xmax

(16)

Note that there seems to be an incoherence between the last three
constraints and the Gauss-Markov model which assumes that there
is a dependency between model parameters and each single observation. Here, the width xmax of the floor is used as observation
twice. We can, however, argue that in fact we have two observations, which have the same value only under the assumption that
the shape of the floorplan is rectangular. Since we will assume a
very low variance for xmax , duplicating this value will be of no

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
306

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

harm for the parameter estimation. Table 1 summarizes the used
model parameters, observations and pseudo-observations as well
as the used abbreviations accordingly.
Each pseudo-observation gives one linear constraint which directly relates the model parameter to the mean µk of the kth
selected component of the Gaussian mixture for the respective
parameter. For room width for example we get:
widthu = µwidth
.
k

(17)

model
parameters

The constraints for the room areas are defined as follows:
widthi × depthi = areai ,

(18)

these constraints have a bi-linear form. Non-linearity is handled
by Taylorization using partial derivatives. This leads to the linear
constraints of the form:
depthi widthi + widthi depthi = ∆areai .

(19)

Note that in this equation the underlined terms are coefficients
in the design matrix while the other terms refer to the respective
model parameters. The coefficients stem from a starting vector
x0 . The selection of a good start vector is important. To derive this vector there are several options. As a first choice prior
knowledge, i.e. pseudo-observations could be used. An alternative would be the parameter values of the topological model.
We found that a better choice is the estimation of these parameters using a relaxed Gauss-Markov model with a design matrix
where the non-linear constraints are omitted. This approach derives good estimations for model parameters up to the width of
the upper and lower rooms and the floor widths. Note that – beside the bi-linear constraints for the areas – these parameters are
related by one single constraint (cf. Equation 14). For this reason, the observed room area of the ith room and the estimated
width widthi is used to derive the two medians of the widths of
the aligned lower and upper rooms:


areai
depthi = median( i |
).
widthi

(20)

Stochastic model:
Observations are assumed to be uncorrelated random variables
with normal distributions. The following standard deviations σ
are assumed:
• embrasures: 0.1 m
• room areas: 0.1 m2
• xmax and ymax : 0.01 m
• interior and exterior walls: 0.03 m
For the pseudo-observations of room depths and widths, standard
deviations of 1 m are assumed. It was noted that the difference
between pseudo-observations and assumed model parameters exceed the standard deviations.
Parameter estimation:
In comparison to the models derived by Loch-Dehbi et al. (2017),
the MAP-estimation includes both prior knowledge in the form of
probability density functions of model parameters and observations such as window locations and room areas. Furthermore, the
representation and exploitation of symmetries enabling a more

observations

pseudoobservations

widthli
widthui
depthli
depthui
widthcorr
depthcorr
waOut
waIn
depthl
depthu

lower room width
upper room width
lower room depth
upper room depth
corridor width
corridor depth
outer wall width
inner wall width
lower depth of rooms
upper depth of rooms

wiDistli,1
wiDistli,2
wiDistui,1
wiDistui,2

lower dist(left embrasure,wall)
lower dist(right embrasure,wall)
upper dist(left embrasure,wall)
upper dist(right embrasure,wall)

wiLocli,1
wiLocli,2
wiLocui,1
wiLocui,2
areai
areacorr
xmax
ymax
µwidth
k
corr
µwidth
k
depth
µk
corr
µdepth
k
waIn
µk
µwaOut
k

left lower embrasure
right lower embrasure
left upper embrasure
right upper embrasure
room area
corridor area
floor width
floor depth
mean of the kth Gaussian
mixture component of a
parameter

Table 1: Overview of the domain parameters, observations and
pseudo-observations for floorplan model selection. The index i
refers to the ith room.
accurate estimation is taken into account. The influence of the
occurrence of repetitive patterns and structures such as the same
window distances to the next wall or recurrent room width has
been investigated in a case study. We performed a simulation using Gauss-Markov modelling and taking these aspects into consideration and were able to compare the impact of such patterns
on the resulting estimations.
All in all, the mentioned approach contributed to the improvement of the accuracy of the estimated model parameters. Table 2 summarizes the achieved accuracies under different simulated conditions. For the estimation of room widths, for instance,
an average accuracy of about 1.32 cm is achieved using a bi-linear
model and assuming that the distances between the window embrasures and the next left or right wall are fix. In this case, the
estimated and the ground truth model are almost identical as depicted on the right of Figure 3. In contrast, we can state worse accuracies without taking repetitive structures such as room widths
into account using a linear model. So far we have assumed that
adjacent rooms are aligned along the corridor (i.e. they have the
same width) and inner and outer walls respectively have the same
widths. In reality in many cases indoor models reveal much more
symmetries. In this paper, we focus on two translational symmetries:
• distances between walls and neighbouring window embrasures are equal
• the widths of rooms are either equal or equal up to a small
integer factor (two or three).
Interestingly, parameter estimation derived so far provides a good
basis for the identification of such symmetries. The covariance

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
307

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

linear model
bilinear variable window distances
bilinear fixed window distances
bilinear room symmetries

room width
68.2
5.75
1.32
0.41

room depth
62.73
1.2
1.14
1.1

inner wall
2.33
1.66
1.36
1.65

outer wall
2.43
1.83
1.51
1.82

corridor width
4.9
3.73
3.14
3.72

corridor depth
63.46
0.96
0.86
0.96

Table 2: Average deviations between the estimated shape parameters of rooms, walls and corridors and the true parameters in [cm].

Figure 3: Floorplan estimation using our approach. Model selection is performed based on linear (top) and bi-linear (bottom) constraints. Symmetries and repetitive patterns (fixed room widths, fixed window distances) are considered. The zoomed part on the right
shows that the estimated (orange) and the ground truth model (black) are almost identical
matrix of the estimated model is given by:
−1
Σxbxb = (AT Σll
A)−1 .

(21)

Its diagonal gives an estimation of the accuracy of the parameter
estimation. Comparison of these accuracies with the differences
of related model parameters such as room widths or distances between walls and embrasures provides hints on symmetries. One
could now use a classical hypothesis test to check for identity. Instead, we have represented symmetries in our models explicitly
and made model comparisons with AIC or BIC as mentioned in
Section 3 (cf. Equations 7 and 9).
Let n be the number of rooms. If symmetries w.r.t distances between walls and embrasures are explicitly represented, the 2n
parameters for distances are replaced by one single model parameter, and the number of columns of the design matrix is reduced accordingly. Please note that the number of observations
and thus the number of rows of the design matrix is not changed.
Thus, in principle the respective likelihoods (cf. Equation 4) can
be compared and used for model selection. The complexities of
the respective models are however different, reflected by a different number of model parameters. That is the reason why penalty
terms as in the Akaike information criterion (Equation 7) or the
Bayesian information criterion (Equation 9) have been applied.
In our case both criteria led to the same results. To assess to what
extent one model outperforms the other normalized likelihoods
(pseudo-probabilities) were used:
LH(i)
LHn (i) = Pn
.
j=1 LH(j)

(22)

In order to represent identity of room widths (up to a factor k) in
the model and the corresponding design matrix explicitly, again
the n parameters for room widths are replaced by one single
model parameter. In the equations the term widthi is replaced

by ki × width where ki is integer depending of the specific ith
room. Again the number of columns of the design matrix is reduced considerably. But since the prior knowledge of this model
parameter is represented by pseudo-observations, the number of
rows is reduced as well. Thus, the likelihoods given by Equation
4 are not comparable immediately. To make them comparable
the covariance matrix Σll is projected to a sub-matrix Σ′ll where
the rows for the pseudo-observations for room widths are omitted. Differences in model complexities are again reflected in the
penalty terms of AIC and BIC, and normalized likelihoods are
derived.
Figure 3 shows the results of the floorplan estimation using our
approach. As mentioned, the model estimation and selection
is performed based on linear (first row) and bi-linear (second
row) constraints. The impact of symmetries such as recurrent
room widths or repetitive window distances is investigated. The
zoomed part on the right reveals that the estimated (orange colour)
and the ground truth model (black colour) are almost identical.
We have also studied how deficiencies of the topological model
can be identified and corrected by the stochastic parameter estimation. Figure 4 illustrates a situation of the floorplan of our
institute in a Wilhelminian style building where a small room (in
orange) of 1.8 m2 is contained in another room. This violates
the assumption that rooms are rectangular – the outer one is not.
Let us assume that the combinatorial reasoning deliberately excludes rooms with an area under a given threshold (e.g. 2 m2 )
and leaves the problem to the stochastic reasoner. The stochastic reasoner as described above starts with an observational error
for the area of that specific room. Having in mind that one small
room of a known size has been omitted, the task is now to identify
the room which is affected and to correct the observed area by the
sum of the areas of the enclosed room and the enclosing room. If
this leads to a model with a considerably larger likelihood, it can
be inferred that the smaller room is enclosed in or adjacent to the
other (although still unclear exactly how). We have studied such

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
308

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

scenarios and used the covariance matrix of the fitted observations:
Σblbl = (AΣxbxb A)T ,
(23)
which has often been used for the identification of outliers in the
observations. Surprisingly this did not lead to satisfactory results.
Instead, we applied the same procedure as given above and used
respective likelihood for model comparison in order to identify
the affected room and to correct the area observation accordingly.

6 EXPERIMENTAL RESULTS
Our approach starts with defining the functional model and a start
vector x̃ of true model parameters according to Figure 2 and the
stochastic model as described above. The stochastic model together with the assumption of normal distributions was used to
generate observational errors ṽ in order to generate noisy observations ˜
l = l + ṽ. We performed this for 1000 test cases and
derived estimated model parameters x
b for the different scenarios
described above. We compared the estimated values x
b with the
true values x̃. The average deviations are listed in Table 2. Obviously, the transition from the linear in the first row to the bi-linear
model in the second row leads to a considerable improvement of
the estimation of room width and room depth. That, however,
is not really surprising since room area is not considered in the
linear model. However, the improvement of model accuracy by
the introduction of further model assumptions is surprising. If we
leave aside corridor width for a moment, both the inclusion of
symmetries of window distances and of room width symmetries
leads to accuracies in the range of 1 cm. This should be compared
to observational accuracy of window locations of 10 cm which illustrates the important role of symmetries for model parameter
estimation.
To get a different impression of this fact, one can look at the redundancies of the model. Let us again assume n rooms and – to
simplify discussion – the same number of windows. If we omit
the window locations for the moment and restrict to the essential model parameters, we have n + 6 model parameters (n room
widths, 2 depths, 1 corridor width and 1 corridor depth and 2
walls) and n + 5 observations (n room areas, 1 corridor area,
3 × xmax and 1 × ymax ). This means that we have an underconstrained system. If we introduce window locations we get 2n additional observation and 2n additional model parameters, which
leads again to an underconstrained system. To achieve an overconstrained system with a certain amount of redundancy, pseudoobservations have been used. Whereas the number of model parameters remains the same, we now get 2n + 8 observations.
This means, however, that prior knowledge on the distribution
of model parameters plays a dominant role. This role is considerably reduced by the explicit representation of symmetries. Window distances lead to a reduction of 2n−1, while room widths
contribute to a reduction of n − 1. Interestingly, both cases lead
to similar results with regard to accuracies.
We also studied how sure we can be with regard to the assumption
of symmetries. As described above, we used AIC and BIC and
derived normalized likelihoods. In all cases, we found that with
0.99 versus 0.01 or better the true symmetric model outperformed
the model without this assumption considerably.
In the current setting, we assume that except for corridors each
room is associated to at least one window. More general settings allow for having rooms without windows. As yet, our approach described how to derive floorplans from sparse observations. The 2D models can be extended following the ideas presented by Loch-Dehbi et al. (2017). Especially door shapes and

Figure 4: Detection and correction of potential errors in the topological model. The small room omitted from the topological
models is estimated using stochastic reasoning.
positions are predicted based on probability density functions approximated by Gaussian mixtures. Prior knowledge about windows is exploited to localize the floor within the considered building. As described by Dehbi et al. (2016), the height of each storey
is derived by kernel density estimations based on 3D point clouds
of façades stemming from an unmanned aerial vehicle. These
conditional predictions depend on a given building style, otherwise the latter can be inferred following the ideas of Henn et al.
(2012).

7 CONCLUSION
This paper presented an approach for predicting and reconstructing a-priori unknown structures in building interiors. In contrast to common approaches, the presented work does not rely on
dense observations such as 3D point clouds. For the prediction of
a floorplan, the areas of the rooms, the footprint and the location
of windows together with a prior knowledge in form of probability density functions is sufficient to estimate model parameters
with high accuracy.
Symmetries and other regularities in man-made objects allow for
a top-down process considering an indoor model with strong constraints and regularities. The latter is however characterized by
linear and bi-linear relations with discrete and continuous parameters. An extensive analysis of a ground truth database yielded a
profound prior knowledge that supported the estimation of floorplans. This includes distributions of model parameters that are
characterized by probability density functions and approximated
by Gaussian mixtures.
In this paper, we focused on the improvement of model parameters in an a-priori estimated topological model that was generated by the use of Constraint Logic Programming. Our approach
uses a Gauss-Markov model and incorporates not only probability density functions, but also observations such as window locations or room areas. It further benefits from the representation of
symmetries in the known exterior model. Bayesian model selection is based on the information criteria AIC and BIC and incorporates observations in order to choose the hypotheses that best
fit the scenario. Errors in the preliminary topological model are
detected and corrected. Thus, accuracies in the range of 1 cm are
possible. The presented approach provides a ranked set of model
hypotheses with corresponding normalized likelihoods.
In this paper, we assume that rooms as well as the corresponding
footprint have a rectangular shape following a Manhattan world
assumption. More general room layouts such as L, T or even
U shaped rooms will be subject of future work. The considered
floorplans characterize office buildings. Model selection for other
building types like housing, where room numbers are not given,
will have to use slightly different models. The derivation of the

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
309

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-2/W4, 2017
ISPRS Geospatial Week 2017, 18–22 September 2017, Wuhan, China

functional model from the topological model will be addressed in
a subsequent article. Our approach does not rely on dense observations such as 3D point clouds from laserscans or range cameras.
However, additional observations of model parameters stemming
from such sensors can be integrated easily. They will improve the
accuracy of the estimated hypothesis.

Loch-Dehbi, S., Dehbi, Y. and Plümer, L., 2017. Estimation of 3d
indoor models with constraint propagation and stochastic reasoning in the absence of indoor measurements. ISPRS International
Journal of Geo-Information.
McLachlan, G. and Peel, D., 2000. Finite Mixture Models. Wiley series in probability and statistics: Applied probability and
statistics, Wiley.

ACKNOWLEDGEMENTS
Niemeier, W., 2008.
Ausgleichungsrechnung:
Auswertemethoden (German). Walter de Gruyter.
We thank Jan-Henrik Haunert for his valuable discussion. The
authors are grateful to Stefan Teutsch for his assistance in preparing the illustrations.

REFERENCES
Akaike, H., 1974. A new look at the statistical model identification. IEEE Transactions on automatic control 19(6), pp. 716–
723.
Becker, S., Peter, M. and Fritsch, D., 2015. Grammar-supported
3d indoor reconstruction from point clouds for ”as-built” BIM.
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences 2(3), pp. 17.

Statistische

Ochmann, S., Vock, R., Wessel, R. and Klein, R., 2016. Automatic reconstruction of parametric building models from indoor
point clouds. Computers & Graphics 54, pp. 94–103.
Pintore, G., Garro, V., Ganovelli, F., Gobbetti, E. and Agus, M.,
2016. Omnidirectional image capture on mobile devices for fast
automatic generation of 2.5d indoor maps. In: 2016 IEEE Winter
Conference on Applications of Computer Vision (WACV), pp. 1–
9.
Rosser, J. F., Smith, G. and Morley, J. G., 2017. Data-driven
estimation of building interior plans. International Journal of
Geographical Information Science pp. 1–23.

Bishop, C., 2007. Pattern recognition and machine learning (information science and statistics). Springer, New York.

Rosser, J., Morley, J. and Smith, G., 2015. Modelling of building interiors with mobile phone sensor data. ISPRS International
Journal of Geo-Information 4(2), pp. 989–1012.

Charman, P., 1994. A constraint-based approach for the generation of floor plans. In: ICTAI, pp. 555–561.

Schwarz, G. et al., 1978. Estimating the dimension of a model.
The annals of statistics 6(2), pp. 461–464.

Dechter, R., 2003. Constraint processing. Elsevier Morgan Kaufmann.

Turner, E. and Zakhor, A., 2014. Floor plan generation and room
labeling of indoor environments from laser range data. In: Computer Graphics Theory and Applications (GRAPP), 2014 International Conference on, IEEE, pp. 1–12.

Dehbi, Y., Staat, C., Mandtler, L. and Plümer, L., 2016. Incremental refinement of facade models with attribute grammar from
3d point clouds. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences III-3, pp. 311–316.
Diakité, A. A. and Zlatanova, S., 2016. First experiments with the
tango tablet for indoor scanning. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences III-4,
pp. 67–72.

Yue, K., Krishnamurti, R. and Grobler, F., 2012. Estimating the
interior layout of buildings using a shape grammar to capture
building style. Journal of Computing in Civil Engineering 26(1),
pp. 113–130.

Förstner, W. and Wrobel, B., 2016. Photogrammetric Computer
Vision: Geometry, Orientation and Reconstruction. Geometry
and Computing, Springer International Publishing.
Frühwirth, T. and Abdennadher, S., 2003. Essentials of constraint
programming. Springer Science & Business Media.
Gröger, G. and Plümer, L., 2012. CityGML–interoperable semantic 3d city models. ISPRS Journal of Photogrammetry and
Remote Sensing 71, pp. 12–33.
Gunduz, M., Isikdag, U. and Basaraner, M., 2016. A review of
recent research in indoor modelling & mapping. ISPRS - International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences XLI-B4, pp. 289–294.
Henn, A., Römer, C., Gröger, G. and Plümer, L., 2012. Automatic classification of building types in 3d city models - using
svms for semantic enrichment of low resolution building data.
Geoinformatica 16(2), pp. 281–306.
Khoshelham, K. and Dı́az-Vilariño, L., 2014. 3d modelling of
interior spaces: Learning the language of indoor architecture. The
International Archives of Photogrammetry, Remote Sensing and
Spatial Information Sciences 40(5), pp. 321.

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-IV-2-W4-303-2017 | © Authors 2017. CC BY 4.0 License.
310